Summary: On July 30, 2026, Chinese startup XinVision released the world's first in-memory computing chip for intelligent vision, NeuroVision-1, achieving 100x energy efficiency over traditional solutions. Breaking the von Neumann bottleneck, it enables pixel-level real-time processing, revolutionizing industrial inspection and autonomous driving. This article analyzes its technology, industry impact, and potential opportunities for Vietnam's manufacturing upgrade.
July 30, 2026, Shanghai — At today's World AI Chip Summit, Chinese AI chip startup XinVision officially launched the world's first in-memory computing chip for intelligent vision applications, the NeuroVision-1. According to official data, the chip achieves an energy efficiency of 0.6 TOPS/W (trillions of operations per watt) in typical vision tasks, over 100x improvement compared to traditional von Neumann architecture chips, marking a new era of integrated storage and computing for AI vision processing.
Breaking the Von Neumann Bottleneck: How In-Memory Computing Works
In traditional chips, data frequently shuttles between processor and memory, consuming significant energy and limiting speed — the so-called "von Neumann bottleneck." XinVision CEO Chen Lifeng stated at the launch: "Visual data is high-bandwidth and high-concurrency; traditional architectures struggle to achieve real-time processing at low power. NeuroVision-1 employs RRAM (resistive random-access memory) based in-memory computing units that embed computation directly into the memory array, enabling core operations like matrix multiplication without moving data, reducing power consumption to below 1% of traditional solutions."
NeuroVision-1 integrates 1,024 in-memory computing cores, each with a 256x256 RRAM crossbar array, capable of parallel large-scale convolution operations. The chip also includes a RISC-V control core and a dedicated vision accelerator, supporting mainstream neural network frameworks (e.g., TensorFlow Lite, PyTorch Mobile). In the MobileNetV3 image classification benchmark, NeuroVision-1 achieves 185 fps at 2.3W, while traditional GPU solutions require over 15W for comparable performance.
Industrial Inspection: First Application Scenario
At the launch, XinVision announced a collaboration with VinSmart, a subsidiary of Vietnamese electronics giant Vingroup, to deploy NeuroVision-1 in automated optical inspection (AOI) for smartphone cameras. A VinSmart factory manager said: "Traditional AOI systems rely on high-performance GPU servers, taking seconds per inspection with high power consumption and heat dissipation issues. With NeuroVision-1, we achieve millisecond defect detection on the production line edge, improving overall inspection efficiency by 10x while reducing power consumption by 90%." Additionally, XinVision partnered with Singaporean logistics company Ninja Van to trial the chip for real-time visual navigation in package sorting.
Vietnam, a key global electronics manufacturing hub, is accelerating automation and smart transformation. According to 2025 data from Vietnam's Ministry of Industry and Trade, electronics manufacturing accounts for over 12% of GDP, but high-end inspection equipment is heavily imported. XinVision co-founder and COO Li Min noted: "Vietnam's market has huge demand for low-power, cost-effective vision chips, especially for small and medium factories that cannot afford expensive GPU clusters. The NeuroVision-1 is priced at only $89, aiming to democratize AI vision capabilities across every production line."
Industry Impact: Intensified Competition in Edge AI Chips
The launch of NeuroVision-1 further heats up the edge AI vision chip market, which already includes NVIDIA's Jetson series, Google's Coral TPU, Huawei's Ascend series, and numerous startups like Groq and Tenstorrent. Analysts believe in-memory computing offers outstanding energy efficiency but faces maturity challenges: RRAM device yield and endurance need improvement, software toolchains are incomplete, and ecosystem compatibility remains to be verified.
XinVision plans a next-generation product using advanced 3D stacking to integrate in-memory computing units with photosensors, achieving a true "sensing-memory-computing integrated" vision chip, with tape-out expected in 2027. The company has completed a Series B+ round led by Sequoia China, Yuanhe Puhua, and Vietnam's sovereign wealth fund SCIC, totaling over $500 million.
Future Outlook: From Chip to System
With the surge of IoT devices, the model of uploading all data to the cloud for processing faces bandwidth, latency, and privacy bottlenecks. Gartner predicts that by 2027, 60% of global AI inference will occur at the edge. XinVision's in-memory computing solution aligns with this trend. Market research firm IDC analyst Wang Zheng noted: "In-memory computing is a key path in the post-Moore era, with disruptive potential in high-throughput tasks like vision. But to become mainstream, it must address volume production and ecosystem adaptation."
For Vietnam, the collaboration with XinVision not only introduces new chip technology but also fosters the local semiconductor ecosystem. The Vietnamese government recently approved the "National Semiconductor Development Strategy 2030," planning to invest $10 billion in chip design centers and packaging/test bases. VinSmart, as Vietnam's largest electronics manufacturer, is actively building a full-chain capability from design to manufacturing. This partnership could accelerate Vietnam's catch-up in AI chips.
At the conclusion, Chen Lifeng revealed that XinVision will set up offices in Hanoi and Ho Chi Minh City, and plans to establish in-memory computing research centers with local universities. "We aim to work with Vietnamese industry partners to drive intelligent vision applications, empowering every corner from factories to cities," he said.
